Online Set Cover: Models & Algorithms

• Online Set Cover is a dynamic covering problem where sets must be irrevocably selected as elements arrive to ensure immediate coverage while minimizing cost.
• Researchers explore diverse models including adversarial, random-order, geometric, subset-arrival, repetition, and dynamic pricing to tackle the problem’s variants.
• Competitive methods utilize logarithmic approximations and innovative rounding techniques to achieve performance guarantees that adapt to instance-specific parameters.

Online set cover is the problem of maintaining an increasing family of active or purchased sets as requests arrive online, so that every arrived element is covered immediately, while minimizing the final cost relative to an optimal offline cover. In the classical weighted model there are $m$ sets $S_1,\dots,S_m$ over a universe of $n$ elements, set $S_i$ has positive cost $c_i$, arrivals are irrevocable, and the maintained family satisfies $$\varnothing=A_0\subseteq A_1\subseteq\cdots\subseteq A_n\subseteq[m]$$ with $j\in\bigcup_{i\in A_j}S_i$ after request $j$. Across recent work, the subject includes adversarial and random-order arrivals, geometric range spaces, subset-arrival and batched models, convex and submodular objectives, dynamic pricing, and fully dynamic maintenance with update-time or recourse guarantees (Kesselheim et al., 25 Aug 2025, Gupta et al., 2021, Khan et al., 2023).

1. Classical formulation and competitive framework

In the standard online set cover model, the set system is fixed in advance, elements arrive one by one, and decisions are irrevocable. Feasibility requires that after each arrival $j$, the current chosen family covers $j$, while the classical weighted objective is

$S_1,\dots,S_m$0

Competitiveness is measured against the optimal offline solution for the final realized instance. In geometric formulations the same ratio is written as

$S_1,\dots,S_m$1

where $S_1,\dots,S_m$2 is the selected family and $S_1,\dots,S_m$3 is an optimal offline cover for all revealed points (Kesselheim et al., 25 Aug 2025, Khan et al., 2023).

Two distinctions organize much of the literature. First, the arrival model may concern elements, subsets, batches, or geometric objects. Second, the information structure may be adversarial, random-order, stochastic, prophet, or sample-based. In the subset-arrival rounding model, for example, only the ground element set $S_1,\dots,S_m$4 is known initially; subset vertices $S_1,\dots,S_m$5 arrive one by one, and on arrival the algorithm learns $S_1,\dots,S_m$6, $S_1,\dots,S_m$7, and $S_1,\dots,S_m$8, and must irrevocably decide whether to select $S_1,\dots,S_m$9 (Byrka et al., 17 Jul 2025). This is a different online interface from classical element-arrival set cover, but it is still a set-cover rounding problem.

The literature also uses several structural parameters. The maximum subset size is

$n$0

the maximum frequency can be written as

$n$1

and in geometric settings one often measures complexity through the number $n$2 of candidate points or the grid scale $n$3 (Byrka et al., 17 Jul 2025, Bender et al., 2024, Khan et al., 2023). Which parameter governs the best guarantee depends strongly on the model.

2. Classical guarantees, random order, and learnable policy classes

For the classical weighted problem, an integral randomized online algorithm with competitive ratio

$n$4

is available, and the paper presenting integral convex-objective algorithms emphasizes that this matches the known optimal dependence, up to constants, for weighted Online Set Cover (Kesselheim et al., 25 Aug 2025). The same logarithmic structure reappears when online set cover is viewed as a special case of online submodular cover: with $n$5 denoting the number of sets and $n$6 the number of arriving elements, the general framework yields

$n$7

for the set cover specialization (Gupta et al., 10 Oct 2025).

A major separation appears in the random-order model. When the elements of the final universe are revealed in a uniformly random permutation, the algorithm "LearnOrCover" achieves expected competitive ratio

$n$8

thereby “circumventing the $n$9 lower bound known in adversarial order” (Gupta et al., 2021). The algorithm maintains a coarse fractional solution that is neither feasible nor monotone increasing, but can nevertheless be rounded online in the random-order model. This replaces the classical monotone-feasible fractional state by a learn-or-cover state driven by a KL-divergence term and a residual-cover term (Gupta et al., 2021).

A different line studies parameterized policy classes rather than a single fixed algorithm. In the class $S_i$0, the algorithm initializes

$S_i$1

maintains the potential

$S_i$2

and, on an uncovered arrival, chooses the minimum $S_i$3 such that $S_i$4, updates $S_i$5 for $S_i$6, and selects at most $S_i$7 subsets from $S_i$8 so that the potential does not increase. Its competitive ratio is

$S_i$9

recovering the classical $c_i$0 bound $c_i$1 and making the policy parameter $c_i$2 learnable from historical instances without leaving a provably competitive class (Zeynali et al., 2020).

Taken together, these results show that the core $c_i$3 adversarial landscape is no longer the only relevant baseline. Random-order information can almost collapse the gap to offline, and policy-class design can expose structural instance dependence without abandoning worst-case guarantees.

3. Geometric online set cover

Geometric online set cover departs from arbitrary set systems by exploiting geometry in the fixed family of sets. A particularly sharp result is for axis-parallel squares in the plane. In this model the family $c_i$4 of squares is known offline from the beginning, points arrive online one by one, and when a point $c_i$5 arrives the algorithm must immediately and irrevocably add squares so that all seen points are covered. For axis-parallel squares of arbitrary sizes, there is a deterministic

$c_i$6

-competitive online algorithm when arrivals come from a known candidate set $c_i$7 of size $c_i$8, and this is tight because any deterministic or randomized online algorithm for set cover for unit squares has competitive ratio $c_i$9 (Khan et al., 2023).

The upper bound is obtained from a monotone offline approximation procedure. In the quadtree formulation one proves monotonicity,

$$\varnothing=A_0\subseteq A_1\subseteq\cdots\subseteq A_n\subseteq[m]$$0

and an offline bound

$$\varnothing=A_0\subseteq A_1\subseteq\cdots\subseteq A_n\subseteq[m]$$1

The online algorithm then simply maintains the current offline solution; monotonicity guarantees that recomputation only adds squares, never deletes them. To replace the coordinate dependence $$\varnothing=A_0\subseteq A_1\subseteq\cdots\subseteq A_n\subseteq[m]$$2 by $$\varnothing=A_0\subseteq A_1\subseteq\cdots\subseteq A_n\subseteq[m]$$3, the paper uses a balanced box decomposition tree of depth $$\varnothing=A_0\subseteq A_1\subseteq\cdots\subseteq A_n\subseteq[m]$$4 (Khan et al., 2023). A plausible implication is that monotone offline structure can be more useful than generic online set-cover reductions when the geometry permits it.

A different geometric line studies translated copies of unit disks and regular unit $$\varnothing=A_0\subseteq A_1\subseteq\cdots\subseteq A_n\subseteq[m]$$5-gons in $$\varnothing=A_0\subseteq A_1\subseteq\cdots\subseteq A_n\subseteq[m]$$6. In the stronger Model-II, only the point set $$\varnothing=A_0\subseteq A_1\subseteq\cdots\subseteq A_n\subseteq[m]$$7 is known in advance, while the geometric objects are not known beforehand and arrive online. Through explicit duality, the hitting-set results imply that the equivalent geometric set cover problem in Model-II admits a deterministic

$$\varnothing=A_0\subseteq A_1\subseteq\cdots\subseteq A_n\subseteq[m]$$8

competitive algorithm for points and translated copies of a disk or a regular $$\varnothing=A_0\subseteq A_1\subseteq\cdots\subseteq A_n\subseteq[m]$$9-gon with $j\in\bigcup_{i\in A_j}S_i$0, while Model-I lower bounds give $j\in\bigcup_{i\in A_j}S_i$1 (De et al., 2023). In particular, regular unit $j\in\bigcup_{i\in A_j}S_i$2-gons recover the unit-square case.

There are also exclusion-constrained geometric variants. In the online class cover problem, a red point set $j\in\bigcup_{i\in A_j}S_i$3 is known in advance, blue points arrive online, and every chosen unit square must be $j\in\bigcup_{i\in A_j}S_i$4-empty. For axis-parallel unit squares in $j\in\bigcup_{i\in A_j}S_i$5, every deterministic online algorithm has competitive ratio at least $j\in\bigcup_{i\in A_j}S_i$6, where $j\in\bigcup_{i\in A_j}S_i$7, and there is a deterministic algorithm with competitive ratio $j\in\bigcup_{i\in A_j}S_i$8 for $j\in\bigcup_{i\in A_j}S_i$9 (De et al., 2023). This is a geometric online set cover with forbidden-region constraints rather than a standard set system.

4. Arrival-model variants: batches, subset arrivals, and repetitions

Several extensions change what arrives online rather than what objective is optimized. In the batched set cover problem, elements arrive in batches $j$0 instead of one by one. Online set cover is the special case where every batch is a singleton. For fractional batched set cover, if the adversary is required to produce batches of VC-dimension at least $j$1, there is a tight lower bound

$j$2

with a matching upper bound $j$3 under the same restriction (Mori et al., 2018). The paper also studies a dedicated simultaneous-update batched primal-dual rule, motivated by the “rich information encoded in the complex interactions between the elements of a batch and the sets that contain them,” although the analysis remains $j$4-competitive in the same asymptotic sense as sequentialized processing (Mori et al., 2018).

Online rounding under subset arrivals isolates a different interface. Here only $j$5 is known in advance; subsets arrive with their neighborhoods, costs, and LP values, and the rounding algorithm must decide immediately whether to select them. Under this model there is an

$j$6

-competitive rounding scheme, where $j$7 is the maximum subset size and is assumed known upfront (Byrka et al., 17 Jul 2025). By contrast, the element-arrival model admits $j$8-competitive rounding, and the previously known subset-arrival guarantee depended on $j$9. The online subset-arrival result therefore closes much of the gap between offline or element-arrival rounding and the stronger subset-arrival model (Byrka et al., 17 Jul 2025).

Online set cover with repetitions changes the covering requirement itself. Elements may arrive multiple times, and if an element $j$0 has appeared $j$1 times so far, the algorithm must ensure that $j$2 is covered by $j$3 different purchased sets. The paper states an

$j$4

-competitive randomized algorithm for the online set cover with repetitions problem and derives it by reduction to admission control to minimize rejections (0803.2842). This is a genuine multi-cover-type strengthening of the standard online model, because repeated requests do not disappear once the element has been covered once.

These variants show that online set cover is not tied to a single arrival grammar. Batch structure, subset-arrival rounding, and repeated-demand coverage each expose different combinatorial bottlenecks.

5. Convex, norm-based, and submodular objectives

The classical weighted objective $j$5 is only one instance of a broader family of online covering objectives. In the convex-objective framework, Online Set Cover is modeled as a special case of Online Generalized Scheduling. Set $j$6 becomes a machine, element $j$7 becomes a job, the processing value is

$j$8

the inner norm is $j$9, and the outer monotone convex function $j$0 aggregates the machine loads. In direct set-cover form, the objective becomes

$j$1

subject to every arriving element being covered by active sets at all times (Kesselheim et al., 25 Aug 2025).

For classical Online Set Cover this framework recovers an integral randomized

$j$2

-competitive algorithm. For a $j$3-bounded convex cost function $j$4, it yields

$j$5

and for an $j$6-norm over multiple linear cost functions it yields

$j$7

For sums or compositions of symmetric norms, the paper gives integral online guarantees such as $j$8 in the $j$9 case and $S_1,\dots,S_m$00 for general symmetric norms (Kesselheim et al., 25 Aug 2025). A central theme is that these are direct integral algorithms, not online rounding procedures applied after solving a convex relaxation.

Online submodular cover generalizes still further. Here the ground set is $S_1,\dots,S_m$01, costs $S_1,\dots,S_m$02 are fixed, and a time-monotone sequence of monotone submodular functions

$S_1,\dots,S_m$03

arrives online. At time $S_1,\dots,S_m$04, the algorithm must output $S_1,\dots,S_m$05 such that

$S_1,\dots,S_m$06

For general online submodular cover the paper proves

$S_1,\dots,S_m$07

and for the $S_1,\dots,S_m$08-increasing subclass it proves

$S_1,\dots,S_m$09

When specialized to online set cover, where the $S_1,\dots,S_m$10 are coverage functions and $S_1,\dots,S_m$11, this becomes exactly

$S_1,\dots,S_m$12

and matches the classical Alon et al. guarantee (Gupta et al., 10 Oct 2025).

A plausible implication is that online set cover now serves as the canonical “covering with irrevocable growth” instance inside a larger theory of integral online convex and submodular covering.

6. Dynamic maintenance, pricing, and stochastic information models

Fully dynamic set cover allows both arrivals and departures of active elements. The maintained family $S_1,\dots,S_m$13 must cover the current active set $S_1,\dots,S_m$14, and competitiveness is measured against the optimum for the current instance. In this model there are algorithms with

$S_1,\dots,S_m$15

and with

$S_1,\dots,S_m$16

In the recourse setting, one can maintain

$S_1,\dots,S_m$17

-competitiveness with constant amortized recourse (Gupta et al., 2016). This differs sharply from classical online set cover because sets may now be both added and removed as the active element set changes.

Dynamic pricing studies implementability rather than approximation alone. In Dynamic Pricing Set Cover, the server can only post surcharges $S_1,\dots,S_m$18 on resources, so the total posted cost is

$S_1,\dots,S_m$19

and an uncovered arriving element $S_1,\dots,S_m$20 chooses

$S_1,\dots,S_m$21

The paper proves the exact characterization

$S_1,\dots,S_m$22

where monotonicity is defined by acyclicity of the induced preference graph, and gives an $S_1,\dots,S_m$23-competitive dynamic pricing algorithm that is optimal for deterministic algorithms (Bender et al., 2024). This links online set cover to posted-price mechanism design without abandoning competitive analysis.

Stochastic, prophet, and sample-based models weaken adversarial uncertainty. In universal stochastic set cover, a map from elements to sets is fixed a priori and the realized subset is random. In the prophet version, the $S_1,\dots,S_m$24 arrivals come from possibly different distributions $S_1,\dots,S_m$25. There is a polynomial-time

$S_1,\dots,S_m$26

-competitive universal algorithm for 1-sample prophet SetCover, improving the earlier $S_1,\dots,S_m$27 stochastic guarantee and requiring only a single sample from each distribution (Gupta et al., 2023). The same reduction also yields an $S_1,\dots,S_m$28-competitive two-stage prophet algorithm and an $S_1,\dots,S_m$29-competitive online-with-a-sample algorithm for the sample fraction $S_1,\dots,S_m$30 (Gupta et al., 2023).

These models broaden the meaning of “online” from pure adversarial arrival to dynamic maintenance, incentive compatibility, and limited distributional information.

7. Problems often conflated with online set cover

A persistent source of confusion is Online Min-Sum Set Cover (MSSC). Despite the name, it is not the standard online set cover problem. In Online MSSC, the algorithm maintains a permutation $S_1,\dots,S_m$31 of a universe $S_1,\dots,S_m$32, a request $S_1,\dots,S_m$33 arrives, the access cost is the position of the first requested element, and the update cost is the Kendall tau distance: $S_1,\dots,S_m$34 The state is a ranking, not a family of purchased sets; the objective is access cost plus reordering cost, not the cost of selected sets (Fotakis et al., 2020, Bienkowski et al., 2022).

This line has its own competitive theory. Against a static benchmark, deterministic online algorithms for the $S_1,\dots,S_m$35-uniform version have lower bound

$S_1,\dots,S_m$36

and upper bound $S_1,\dots,S_m$37, while the efficient memoryless algorithm Move-All-Equally has lower bound $S_1,\dots,S_m$38 against the static optimum and bounds $S_1,\dots,S_m$39 and $S_1,\dots,S_m$40 against the dynamic optimum (Fotakis et al., 2020). Against a dynamic optimum, a later paper gives a computationally efficient randomized $S_1,\dots,S_m$41-competitive algorithm and a deterministic existential $S_1,\dots,S_m$42-competitive algorithm (Bienkowski et al., 2022). In a repeated-decision or online-learning formulation, MSSC appears as Pandora’s Box with values in $S_1,\dots,S_m$43, and one obtains approximate no-regret guarantees rather than classical competitive ratios (Gergatsouli et al., 2022).

Another distinct direction is restricted-information set cover in a static oracle model. “Set Cover in Sub-linear Time” studies a fixed instance under $S_1,\dots,S_m$44 and $S_1,\dots,S_m$45 queries and proves query bounds such as

$S_1,\dots,S_m$46

with matching lower bounds in several regimes, but this is not an online-arrival model (Indyk et al., 2019). The online difficulty there is replaced by query complexity on a static input.

The broader lesson is terminological as much as technical: “online set cover” properly refers to irrevocable covering under arrival uncertainty, whereas min-sum ranking, Pandora-style search, and static sublinear-query set cover are separate problems with different state spaces, cost models, and lower-bound regimes.